A method and system for growing the consciousness of a body-embodied intelligent agent

Abstract

The application provides a body intelligent agent consciousness growth training method and system, which solves the problems of dynamic growth of the body intelligent agent, constructs a multi-dimensional corpus integrating philosophical classics and folk proverbs, trains a situation understanding model, maps multi-modal perception to a standardized situation vector, and then matches and activates a corresponding decision logic unit by relying on a pre-stored decision operator library, generates a comprehensive action plan through meta-rule arbitration when a serious conflict occurs, and finally adopts a reinforcement learning closed loop to continuously optimize the parameters of the decision unit according to the environmental feedback, so that the value-oriented continuous growth of the body intelligent agent is realized. The application has the advantages of good dynamic growth, situation perception and value judgment capability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for training the consciousness growth of an embodied intelligent agent. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially breakthroughs in large language models and multimodal technologies, embodied intelligent agents are increasingly being used in the field of home services. However, existing domestic service robots generally suffer from poor environmental adaptability and insufficient flexibility, making it difficult to achieve deep contextual awareness and autonomous decision-making in complex and ever-changing real-world home scenarios. Specifically, traditional methods typically rely on preset instructions or rules, lacking a dynamic understanding of user intentions, emotional changes, and social interaction signals, resulting in rigid services and an inability to cope with unexpected situations.

[0003] Current mainstream technologies primarily rely on large-scale data training and massive computing power. However, the home environment is highly personalized and inexhaustible, making it difficult for such methods to achieve effective generalization with limited data. Furthermore, once trained, the model's capabilities become fixed, failing to self-optimize through continuous interaction with the environment during service provision, thus hindering its ability to meet long-term personalized needs. In addition, existing models have a weak ability to recognize complex social signals with dual intentions, such as hypocrisy and pretense, making them prone to misjudgment in interpersonal interactions.

[0004] Therefore, a new training method is needed that enables embodied intelligent agents to grow dynamically in real family environments and possess situational awareness and value judgment capabilities, in order to overcome the limitations of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing a method and system for training the embodied intelligent agent's consciousness growth, which is dynamically evolving, possesses situational awareness and value judgment capabilities, and can enhance social interaction security and proactive defense awareness through the Eastern wisdom of observing words and expressions and understanding that appearance reflects the heart.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an embodied intelligent body consciousness growth training system, comprising: This context-based decision operator library stores multiple predefined decision logic units. Each unit encapsulates an evaluation or action mode for a specific context and includes preconditions defining its applicable scenarios. The decision logic units are implemented as executable code modules. Each unit contains a precondition function, a processing logic function, and adjustable parameters. The precondition function calculates the matching degree between the input context feature vector and the unit's applicable scope. The processing logic function takes the context vector as input and outputs an action suggestion vector. Adjustable parameters include confidence weights. Preconditions can be implemented based on vector similarity, rule thresholds, or classification models. The operator library adopts a modular design, supporting dynamic addition, deletion, and parameter updates.

[0007] The context understanding module is configured to receive and process multimodal sensing data, converting it into a standardized context feature vector. Multimodal perception data includes visual images, audio signals, semantic text, and environmental sensor data. The contextual understanding module first extracts features from each modality of data, such as using pre-trained ResNet to extract image features, wav2vec to extract audio features, and BERT to extract text features. Then, it fuses the multimodal features through a multi-head attention mechanism to generate a fixed-dimensional contextual feature vector. This vector represents the semantic and social meaning of the current context in a pre-trained embedding space.

[0008] The facial recognition and analysis module, connected to the context understanding module, is configured to perform facial detection and feature extraction on visual images in multimodal perception data, and generate the user's facial feature vector, trust score, and alertness level based on a pre-trained facial-psychology mapping model. The facial recognition and analysis module includes: a facial feature extraction unit, used to detect faces in images and extract facial features, contour features, texture features, and dynamic micro-expression features; a facial-psychology mapping model, trained based on Eastern physiognomy knowledge and a large amount of labeled data, used to map the extracted facial features to corresponding psychological traits and risk levels; and a multimodal fusion unit, used to fuse facial feature vectors with other modal features output by the contextual understanding module to generate a more comprehensive user intent judgment vector. The output of this module serves as an important component of the contextual feature vector for subsequent decision-making.

[0009] The decision engine module is connected to the context decision operator library and the context understanding module, respectively. It is configured to match and activate at least one candidate decision logic unit that meets its preconditions from the context decision operator library based on the context feature vector, and obtain the action suggestion vector output by each activated candidate decision logic unit. The decision engine calculates the similarity between the context feature vector and the preconditions of each decision logic unit in the operator library, selecting units with similarity exceeding a threshold as candidates. Each activated unit independently runs its processing logic, outputting an action suggestion vector representing the suggested action type, parameters, or direction, and a confidence score representing the unit's certainty of applicability to the current context.

[0010] The training and optimization module is configured to acquire feedback data after the embodied agent performs actions, and adjust the internal parameters or confidence levels of the activated decision-making logic units based on the feedback data. Feedback data includes explicit user feedback, such as voice evaluation and button operations; implicit feedback, such as changes in user facial expressions, task completion time, and changes in environmental state; and preset value functions, such as safety indicators and efficiency indicators. The training and optimization module employs the policy gradient method from reinforcement learning, calculates reward signals based on feedback, and updates the internal parameters or confidence weights of the neural network weights of the units participating in decision-making through backpropagation, achieving continuous online learning.

[0011] Preferably, in the above-mentioned embodied intelligent agent consciousness growth training system, the situational understanding module includes a pre-trained vector encoding model, which is trained based on a corpus containing classical philosophical concepts and folk sayings, wherein the folk sayings are vectorized and labeled in multiple dimensions according to their implied emotional type, social intention and value orientation.

[0012] The vector encoding model employs a Transformer architecture and undergoes contrastive learning pre-training on a constructed multi-source corpus. The corpus contains at least two types of data: first, core concepts and their annotations extracted from classics such as the *Tao Te Ching* and *Zhuangzi*; and second, proverbs and colloquialisms compiled from folk language, manually labeled according to emotional intensity (e.g., anger, joy), social intention (e.g., hypocrisy, sincerity, deception), and value orientation (e.g., selfishness, altruism). The training objective is to shorten the distance between semantically similar texts (such as similar proverbs) in the vector space and widen the distance between texts of different categories, thereby enabling the model to capture the implicit features of complex social signals.

[0013] Preferably, in the above-mentioned embodied intelligent agent consciousness growth training system, the decision engine module further includes a meta-rule arbitration unit, configured to, when the degree of conflict between the action suggestion vectors output by multiple activated candidate decision logic units exceeds a preset threshold, perform a comprehensive analysis of the conflicting action suggestion vectors according to preset meta-rules, and generate a comprehensive action vector as the final output.

[0014] The degree of conflict is measured by calculating the cosine similarity or Euclidean distance between action suggestion vectors. When the minimum similarity is below a threshold, such as -0.5, a serious conflict is considered to exist. The meta-rule arbitration unit executes a four-step algorithm of suspension-lifting-reconstruction-downward: First, the current conflict is suspended, and the conflict vector is input into a high-level value network, which is trained based on a preset value target to generate an abstract value vector; then, the conflict is reconstructed, and the weighted combination of the abstract value vector and each candidate vector is input into a generative adversarial network to generate a comprehensive action vector that takes into account the rationality of all parties; finally, it is downwarded into specific action instructions.

[0015] In the aforementioned embodied intelligent agent consciousness growth training system, the situational decision operator library includes cognitive operators, trade-off operators, action operators, and defense operators; among them, the defense operators are configured to output the confidence level of the consistency between the user's behavioral intention and historical behavioral patterns based on the input surface behavioral characteristics and historical behavioral pattern characteristics representing social interaction.

[0016] The defensive operator incorporates a behavioral consistency assessment model. This model compares the surface features of the current interaction, such as language content, facial expressions, and body language, with historical behavioral pattern vectors extracted from long-term interaction records, calculating a deviation score. When the deviation exceeds a preset threshold, a high-confidence intent anomaly signal is output, and corresponding response strategies are generated, such as maintaining observation, providing a subtle reminder, or initiating a security protocol. This operator does not directly make moral judgments about the user; instead, it identifies behavioral pattern deviations based on a data statistical model.

[0017] In the aforementioned embodied intelligent agent consciousness growth training system, the defensive operators also include a facial defense subunit. This subunit generates preliminary interaction strategies based on the trust score and alertness level output by the facial perception and analysis module. When the risk level mapped by facial features exceeds a preset threshold, this subunit activates corresponding defensive mindset operators, such as the "Facial Features Reflect Mind" operator, outputting strategy suggestions such as increasing alertness, verifying identity, maintaining observation, and cautious interaction. These suggestions are then used as a dimension or confidence adjustment factor in the action suggestion vector for the decision engine to reference.

[0018] In the aforementioned embodied intelligent agent consciousness growth training system, the training and optimization module adopts a reinforcement learning mechanism. Feedback data is used to calculate the value function gap with the preset ideal state, and the parameters of the decision logic unit are adjusted accordingly.

[0019] A method for developing embodied intelligence consciousness includes the following steps: S1: A context-based decision operator library is pre-built, storing multiple decision logic units. Each decision logic unit includes preconditions and processing logic. The construction of the operator library includes unit design, parameter initialization, and unit registration. The processing logic of each unit can be implemented based on rules, shallow models, or neural networks. Preconditions are initially set through offline simulation or expert experience and can be dynamically adjusted during training.

[0020] S2: Acquire multimodal perception data of the agent and convert it into standardized contextual feature vectors through a pre-trained contextual understanding model. The perception data stream is input in real time, and the contextual understanding model uses an end-to-end multimodal encoder to output fixed-dimensional vectors. This model has been fine-tuned on corpora containing social signal annotations during the pre-training stage to effectively encode complex interpersonal interaction features.

[0021] S3: Based on the context feature vector, match and activate one or more candidate decision logic units whose preconditions are met from the context decision operator library. The matching process is achieved by calculating the cosine similarity between the context vector and the precondition vector of each unit, and selecting the top K units with similarity above a threshold as candidates. The number of candidates is dynamically variable, and the value of K is adaptively adjusted according to the context complexity.

[0022] S4: Execute the processing logic of each activated candidate decision logic unit to obtain the corresponding action suggestion; each candidate unit runs its processing logic independently and outputs an action suggestion vector and confidence level. The action suggestion vector represents the type, parameters and expected effect of the action in a high-dimensional space, and can be mapped to specific instructions through the decoder, such as carrying water corresponding to moving to the kitchen - picking up the water cup - filling the water cup - moving to the elderly person - handing over the water.

[0023] S5: Based on preset decision rules, determine and execute the final action instruction from at least one obtained action suggestion; if there is only one candidate, adopt it directly; if there are multiple candidates and no conflict, adopt the weighted average or the one with the highest confidence; if there is a conflict, trigger meta-rule arbitration. The final action instruction is decomposed into underlying control signals by the motion planning module.

[0024] S6: After executing the final action command, acquire environmental feedback and update the relevant parameters of the activated decision logic unit based on the feedback. After collecting the feedback data, the system calculates the difference between the value function of the current state and the ideal state, uses this difference as a loss signal, and updates the parameters of the activated unit through backpropagation. The update process uses small-batch experience replay to stabilize the learning process.

[0025] In the above-mentioned embodied intelligent agent consciousness growth training method, step S2 further includes: before or simultaneously converting multimodal perception data into situational feature vectors, calling the facial perception and analysis module to extract facial features and map mindset on visual images, generating facial feature vectors and their corresponding trust scores and alertness levels, and using this information as components of the situational feature vectors.

[0026] In the aforementioned embodied agent consciousness growth training method, step S5 further includes: when multiple conflicting action suggestions exist and the degree of conflict exceeds a threshold, a meta-rule arbitration process is invoked to comprehensively weigh the conflicting action suggestions and generate a comprehensive final action instruction. The meta-rule arbitration process first inputs the conflict vector into a pre-trained conflict resolution network, which aims to maximize the value function and outputs a comprehensive vector. Subsequently, a differentiable planning layer decodes the comprehensive vector into a specific action sequence, ensuring the executability of the actions.

[0027] In the aforementioned embodied agent consciousness growth training method, the contextual understanding model is trained using a contrastive learning approach based on a multi-source corpus containing philosophical classics and folk proverbs. This allows the model to map different contexts with similar social meanings to neighboring locations in the vector space. The contrastive learning employs the InfoNCE loss function, using similar proverbs (such as multiple idioms expressing hypocrisy) as positive samples and dissimilar proverbs as negative samples. The trained encoder can map contextual text descriptions and visual scenes to the same semantic space, achieving cross-modal social signal alignment.

[0028] In the aforementioned embodied agent consciousness growth training method, step S6, which updates the parameters of the decision-making logic unit based on environmental feedback, specifically includes: calculating the distance change to the preset value goal based on environmental feedback, and using the distance change as a reinforcement learning signal to adjust the internal weights of the relevant decision-making logic unit. The preset value goal consists of user-configurable multi-dimensional goals, such as increased user pleasure and reduced task time. The system quantifies these indicators in real time through emotion recognition and task tracking modules, calculating a weighted sum as the reward. A policy gradient algorithm is used to maximize the expected cumulative reward and update the unit parameters.

[0029] In the aforementioned embodied agent awareness growth training method, the decision logic unit includes a defense unit for identifying hypocritical patterns. The processing logic of this unit is as follows: it analyzes the difference between surface behavioral features and implicit intention features in the input context. When the difference exceeds a preset threshold, it outputs a high-confidence hypocrisy identification result and corresponding defensive action suggestions. The defense unit incorporates a dual-channel neural network: channel one processes surface behavior, including language content and explicit actions; channel two processes implicit signals, including micro-expressions, tone of voice, and contextual inconsistencies. The outputs of the two channels are processed through a difference calculation layer to obtain an inconsistency score. When the score exceeds a threshold, a defense strategy generator is triggered. The strategy generator, based on rules or few-sample learning, outputs suggestions such as remaining vigilant, subtly verifying information, and initiating security protocols. The design of this unit avoids characterizing the user's personality, focusing only on the suspiciousness of behavioral patterns.

[0030] In the aforementioned embodied agent consciousness growth training method, the decision logic unit also includes a defense operator based on the idea that "appearance reflects mind." The precondition for this operator is that the risk level output by the facial perception and analysis module exceeds a preset threshold. Its processing logic is as follows: it acquires facial feature vectors and, based on a facial-mind mapping model, generates an evaluation result containing a trust score and alertness level, thereby outputting corresponding interaction strategy suggestions, such as identity verification, cautious interaction, or normal interaction. The output of this operator serves as a supplement or correction to the action suggestion vector, guiding the agent to take appropriate preventative measures at the initial stage or key nodes of interaction with the user.

[0031] Compared with existing technologies, the advantages of this invention are as follows: By constructing a reusable structured decision operator library, philosophical wisdom is transformed into modular knowledge units. Operator combinations address diverse scenarios, avoiding massive data collection and annotation, and significantly reducing data dependence. The invention innovatively introduces a facial perception and analysis module, transforming the Eastern philosophical idea that "appearance arises from the heart" into a computable model, enabling the intelligent agent to initially judge user intentions and risk levels through vision, thus establishing proactive awareness from the initial interaction stage. A five-step closed-loop training process and reinforcement learning mechanism are adopted, allowing the intelligent agent to update operator parameters online based on user feedback in real-world services, achieving dynamic growth and adapting to the personalized needs of family scenarios. Folk sayings and proverbs are incorporated into the corpus and subjected to double-layer intention annotation. The trained vector encoding model captures deep social characteristics such as apparent goodwill and underlying malice, enabling the intelligent agent to accurately identify abnormal intentions and avoid misjudgments. A meta-rule arbitration unit is set up to generate a comprehensive action plan that takes into account the rationality of all parties when multiple decision suggestions conflict. This enables the agent to make autonomous decisions in value conflicts such as safety and efficiency. A modular operator architecture is adopted, making the decision path traceable. Each step of the logic is clear and auditable, which facilitates developer debugging and user trust. A vector similarity fast matching operator is used, and the decision path length is constant-level. Compared with traditional rule base traversal or large model inference, the response speed is faster and meets the real-time requirements of home scenarios. The operators are derived from philosophy and have cross-scenario universality. The operator library trained in home services can be transferred to scenarios such as medical care and education tutoring with only minor adjustments for reuse, reducing the development cost of new scenarios. Pre-set multi-dimensional value goals as decision guides, combined with the identification mechanism of abnormal intentions by defensive operators, prevents the agent from being maliciously used at the algorithm level, and achieves the inherent alignment of technical solutions with human ethics. By combining facial judgment and behavioral pattern analysis, a complete social perception chain is built for the agent, from observing words to observing expressions, and then to understanding the mind, which significantly improves its security and robustness in complex interpersonal interactions. Attached Figure Description

[0032] Figure 1 This is the overall system architecture diagram of the present invention; Figure 2 This is a schematic diagram of the classification of the mental operator library of the present invention; Figure 3 This is a flowchart of the three-layer pyramid architecture of the present invention; Figure 4 This is a flowchart of the five-step training method of the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Example 1

[0034] like Figure 1-4As shown, this embodiment demonstrates how an intelligent agent can achieve situational awareness and autonomous decision-making through the collaborative work of various system modules in a routine household service. The scenario is set as follows: a housekeeping robot is performing a cleaning task in the living room and detects that an elderly person is coughing, there are water stains on the coffee table, and a child is playing with toys in a corner. The specific implementation steps are as follows: S1: System Initialization First, a contextual decision operator library is loaded, containing ninety-nine predefined decision logic units, with each unit's initial weights set to a uniform distribution. Simultaneously, a multimodal encoder trained on philosophical corpora and folk sayings is loaded as the contextual understanding model. The preset value target vectors are: user satisfaction 0.8, task completion 0.7, and safety index 0.9.

[0035] S2: Multimodal Perception and Contextual Understanding The robot uses a vision module to recognize an elderly person's coughing motion, water stains on a coffee table, and a child's playing posture; an audio module to recognize cough sounds; and environmental sensors to acquire room temperature data. The contextual understanding module inputs this multimodal information into a pre-trained encoder, which, after being fused through a multi-head attention mechanism, generates a high-dimensional feature vector representing the current context.

[0036] S3: Operator Matching and Activation The decision engine calculates the similarity between the context feature vector and the precondition vector of each unit in the operator library, and selects units with a matching degree exceeding a threshold as candidates. The candidate units activated in this instance include the unit for caring for the elderly (high confidence), the unit for clearing obstacles (medium confidence), and the unit for guiding children (slightly low confidence).

[0037] S4: Action Proposal Generation The "Caring for the Elderly" unit executes its processing logic, suggesting that the elderly person receive warm water; the "Clearing Obstacles" unit suggests wiping up water stains first; and the "Guiding Children" unit suggests reminding children to tidy up their toys. Each suggestion is accompanied by a confidence score.

[0038] S5: Conflict Detection and Decision Making The system calculates the degree of conflict among the three action suggestions and finds no obvious contradictions. Therefore, it uses a confidence-weighted average method to generate a comprehensive action plan. This plan is decoded into a specific instruction sequence: the robot first moves to the kitchen to fetch water, saying to the child along the way, "Sweetie, shall we tidy up the toys first?", then hands the water to the elderly person, and finally returns to wipe up the water stains.

[0039] S6: Execution and Feedback After executing the above instructions, the robot observed the elderly person smiling and nodding, and the child starting to tidy up their toys. The training and optimization module calculated the gap between the current state and the preset value target: user satisfaction increased from 0.6 to 0.75, and task completion increased from 0.5 to 0.8, resulting in a positive reward. Based on this reward signal, the system uses a proximal strategy optimization algorithm to update the internal weights of the three units: caring for the elderly, clearing obstacles, and guiding the child, making it more likely to choose this combination of actions in similar future situations. Example 2

[0040] This embodiment demonstrates how a meta-rule arbitration unit generates a creative solution when multiple action suggestions severely conflict. The scenario is as follows: A robot is assisting with cooking in the kitchen when it suddenly detects smoke in the living room, potentially indicating a fire risk, and simultaneously hears an elderly person fall in the bedroom. Both events require an urgent response, but there is only one robot. The specific implementation steps are as follows: S1: Contextual Understanding The robot receives multimodal inputs: smoke alarm signals (audio and sensor data), fall sounds (audio), and vital sign data of the elderly (wearable devices). The contextual understanding module fuses this information and outputs a high-dimensional contextual feature vector representing the level of urgency and safety risk.

[0041] S2: Operator activation The fire emergency unit was activated, suggesting immediately go to the living room to confirm the fire and call the fire department; this suggestion has a high confidence level. The medical emergency unit was also activated, suggesting immediately go to the bedroom to check on the elderly person's condition; this suggestion also has a high confidence level. The two suggestions are completely opposite.

[0042] S3: Collision Detection The system calculates the similarity between the two action suggestion vectors and finds that they are severely negatively correlated, thus determining that they are in serious conflict.

[0043] S4: Meta-rule Arbitration The decision engine inputs conflict suggestions and situational feature vectors into the meta-rule arbitration unit. This unit executes a four-step algorithm: suspension-elevation-reconstruction-downward flow. First, conflict vectors are temporarily stored without direct selection. Then, they are input into a high-level value network, outputting an abstract value vector representing the trade-off between safety priority and life priority. Next, a weighted combination of the abstract value vector and the original suggestions is input into a generative adversarial network to generate a comprehensive action vector. Finally, this vector is decoded into a specific action sequence: The robot first moves to the living room doorway to assess the fire level, visually determining it to be a small area of ​​smoke, not an open flame. Simultaneously, it activates the smoke alarm via the Internet of Things and notifies the property management. Then, it goes to the bedroom to check on the elderly person, finding that they have only slipped and are conscious. Finally, it returns to the living room to deal with the source of the smoke, which is simply a pot that has burned dry in the kitchen.

[0044] S5: Execution and Feedback After the operation, the fire was brought under control, the elderly person was unharmed, and the user's post-incident assessment indicated that the handling was appropriate. The system received a positive reward, updated relevant unit parameters, and fine-tuned the internal network weights of the meta-rule arbitration unit. Example 3

[0045] This embodiment demonstrates how defensive operators can identify complex social intentions that appear benevolent but are actually malicious, thus protecting user safety. The scenario is as follows: A robot is working in the kitchen when a stranger's voice is heard in the living room. The visitor says to the elderly woman, "Grandma, I'm a friend of your son. I've come to visit you and brought a box of health supplements." The elderly woman welcomes the visitor and lets him in. The specific implementation steps are as follows: S1: Contextual Understanding Visual recognition revealed an overly enthusiastic visitor expression; micro-expression analysis showed an upturned mouth but no contraction of the orbicularis oculi muscle, and the health product packaging lacked proper labeling. Speech recognition indicated an exaggerated tone and rapid speech, mentioning "your son" but without specifying his name. Contextual memory analysis revealed recent cases of health product fraud in the community, and no record of this visitor in the elderly person's social media history. The contextual understanding module outputs surface behavioral characteristics, including high friendliness and high-value gifts, as well as implicit signal characteristics, including abnormal micro-expressions, suspicious packaging, and semantic ambiguity.

[0046] S2: Operator activation The consistency evaluation of defensive operators is activated. This operator incorporates a dual-channel neural network: one channel processes surface behavior, and another channel processes implicit signals. The difference calculation layer outputs an inconsistency score, which is set to a preset threshold.

[0047] S3: Action Suggestion Generation The defense operator outputs an evaluation result: the intent is anomalous with a very high confidence level, and generates a response strategy suggestion: maintain observation, verify tactfully, and activate security protocols if necessary. Simultaneously, the hospitality unit and security defense unit are also activated, but the defense operator has the highest confidence level.

[0048] S4: Decision-making and Execution The decision engine adopts the defense operator's suggestion and generates specific action instructions: The robot moves to the living room, smiles gently, and says, "Grandma, let me help you put away your health supplements." It also casually adds, "By the way, your son just called and said he'll be home late, so you should rest first." This statement implies that the son didn't mention the visitor, giving the elderly woman a hint while simultaneously providing an out for the visitor. The robot activates its recording function, storing the recording locally for later analysis and continuous monitoring of visitor behavior.

[0049] S5: Feedback and Learning If the visitor is a scammer, they will appear flustered and leave under some pretext. The robot will then remind the elderly person to be more vigilant and update the community's fraud characteristic database. If the visitor is indeed a friend, the robot will apologize after the explanation and state that it was for safety reasons. In either case, the system will calculate a reward based on the final result, increasing the elderly person's safety index and updating the relevant operator parameters. Example 4

[0050] This embodiment demonstrates how the system learns the personalized preferences of different users through long-term interaction, becoming increasingly user-friendly. The scenario is set as follows: a family with three members; the grandfather prefers quiet and enjoys tea; the grandmother prefers cleanliness and enjoys chatting; and the child is active and enjoys snacks. The robot needs to adapt to the differentiated needs of these users. The specific implementation steps are as follows: S1: User Profile Initialization The system creates a unique user profile vector for each family member, with initial values ​​set to default preferences. As interactions increase, user feedback is recorded after each service session, and the corresponding profile vector is updated accordingly.

[0051] S2: Context Awareness and Personalized Matching One day, Grandpa was sitting on the sofa watching TV, his teacup empty on the coffee table in front of him. The contextual understanding module identified features such as Grandpa, the empty teacup, and watching TV. Combining these with Grandpa's profile—his preference for tea and quiet—it output a contextual vector containing the user ID. The decision engine adjusted the operator activation weights based on the user ID: in the tea-serving unit, the version for Grandpa prioritized brewing green tea with gentle movements; the version for Grandma might include a greeting.

[0052] S3: Long-term learning outcomes After dozens of interactions, the system's accuracy in predicting different users' preferences improved. For example, when it detected that Grandpa was coughing, the system not only brought him warm water but also proactively adjusted the room temperature and played his favorite soft music. Each positive user feedback served as a positive reward signal, reinforcing the corresponding user-context-action connection weight. Example 5

[0053] This embodiment demonstrates the portability of the operator library, enabling rapid adaptation of an operator library trained in home services to a medical care scenario. The scenario is set as follows: deploying an AI agent, already well-trained in home services, to a nursing home to undertake elderly care tasks. The implementation steps are as follows: S1: Corpus Expansion Based on the existing philosophical corpus and folk sayings, we supplemented it with professional terms and common situational descriptions from the medical and nursing fields, such as medication reminders and blood pressure monitoring. The new corpus was vectorized and annotated, and then integrated into the original training set.

[0054] S2: Operator fine-tuning Ninety-nine core decision-making logic units are retained, with adjustments made only to the preconditions of some units for medical scenarios. For example, the unit for caring for the elderly adds recognition weights for preconditions such as medication time and abnormal vital signs. A new medical emergency unit is also added, encapsulating professional processing logic such as CPR and calling for emergency medical services.

[0055] S3: Rapid Deployment The migration process only requires collecting a small amount of sample data in the new scenario, about one-tenth of the original data volume, to fine-tune the model. The prototype of the companion robot, which originally required six months of development, was completed in two weeks, significantly reducing development costs.

[0056] S4: Effect Verification After deployment, the robot can accurately remind the elderly to take their medication on time, recognize emergencies such as falls, and has the ability to chat with the elderly and understand their emotions. Caregivers reported that it's like having a well-behaved housekeeping robot; it learned caregiving tasks with just a little instruction. Example 6

[0057] This embodiment demonstrates how the facial perception and analysis module functions in the initial interaction between an intelligent agent and a stranger, enabling intelligent judgment based on observation of facial expressions and intuition. The scenario is set as a housekeeping robot's first encounter with a visitor who claims to be a friend of an elderly person.

[0058] S1: Situational Understanding and Facial Perception The robot captures visitors' facial images through a vision module. The facial perception and analysis module first performs facial detection, extracting facial features, contour features, and micro-expressions. The visitor's features include: small eyes, shifty gaze, and upturned corners of the mouth but no change in the muscles around the eyes. Based on these features, the facial-psychology mapping model outputs a trust score and risk level, and labels the visitor as potentially dishonest or concealing information.

[0059] S2: Multimodal fusion and operator activation The contextual understanding module simultaneously processes the visitor's speech, including whether the speech is too fast or unclear, as well as their behavior, such as hands in pockets or slightly turned body. This information is then fused with facial features to form a complete contextual feature vector. Based on this vector, the decision engine activates both the standard hospitality unit and the newly added "Facial Features Based on Intent" defense operator.

[0060] S3: Action Recommendation Generation and Decision Making The "Hospitality Unit" outputs a smiling greeting and suggests seating. The "Appearance Reflects Inner Feelings" defense operator outputs "Keep Observing," suggesting a strategy to verify identity, along with a low-confidence correction factor. The decision engine combines both to generate the final interaction command: the robot greets with a smile, saying naturally, "Hello! It's a pleasure to meet you. The elderly often mention you. May I ask your name? Let me register your visit information so the elderly can feel at ease." The robot maintains a safe distance from the visitor and continuously monitors their micro-expressions and body language.

[0061] S4: Feedback and Learning If the visitor is a true friend, they will answer frankly and may appreciate the robot's thoughtfulness. The system receives positive feedback that no security incident occurred, which slightly lowers the activation threshold of the "appearance-based" operator in similar future situations, making it more accurate. If the visitor is a scammer, they may appear flustered and leave under some pretext. The robot records this unusual interaction, updates its risk pattern database, and provides warnings to the elderly person in the future.

[0062] In summary, the principle of this embodiment is as follows: By constructing a structured corpus containing multi-dimensional vectorized annotations extracted from philosophical classics and folk sayings, and training a context understanding model based on this corpus, the multimodal information perceived by the embodied agent is mapped into standardized context feature vectors. Furthermore, relying on a pre-stored operator library containing multiple decision logic units encapsulating specific context evaluation or action patterns, the decision engine matches and activates corresponding candidate units based on the context feature vectors to generate action suggestions. When multiple suggestions seriously conflict, a meta-rule arbitration unit is triggered to comprehensively weigh the options and generate a creative comprehensive action plan. Finally, through a five-step closed-loop training process employing reinforcement learning, the parameters of the activated decision logic units are continuously optimized based on post-action environmental feedback, enabling the agent to achieve value-oriented, interpretable, continuous growth and autonomous decision-making in its interaction with the real environment.

[0063] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0064] Although this paper frequently uses terms such as context decision operator library, context understanding module, decision engine module, and training and optimization module, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A system for training the growth of embodied intelligent consciousness, characterized in that, include: The context decision operator library stores multiple predefined decision logic units. Each decision logic unit encapsulates an evaluation or action mode corresponding to a specific context and includes the preconditions that define its applicable scenarios. The context understanding module is configured to receive and process multimodal sensing data, converting it into a standardized context feature vector. The decision engine module is connected to the context decision operator library and the context understanding module, respectively, and is configured to match and activate at least one candidate decision logic unit that meets its preconditions from the context decision operator library according to the context feature vector, and obtain the action suggestion vector output by each activated candidate decision logic unit. The training and optimization module is configured to acquire feedback data after the embodied agent performs an action, and adjust the internal parameters or confidence level of the activated decision logic unit based on the feedback data.

2. The embodied intelligent agent consciousness growth training system according to claim 1, characterized in that, The context understanding module includes a pre-trained vector encoding model, which is trained on a corpus containing classical philosophical concepts and folk sayings. The folk sayings are vectorized and annotated in multiple dimensions according to their implied emotional type, social intention, and value orientation.

3. The embodied intelligent agent consciousness growth training system according to claim 1, characterized in that, The decision engine module also includes a meta-rule arbitration unit, configured to perform a comprehensive analysis of the conflicting action suggestion vectors according to preset meta-rules when the degree of conflict between the action suggestion vectors output by multiple activated candidate decision logic units exceeds a preset threshold, and generate a comprehensive action vector as the final output.

4. The embodied intelligent agent consciousness growth training system according to claim 1, characterized in that, The contextual decision operator library includes cognitive operators, trade-off operators, action operators, and defensive operators; wherein, the defensive operators are configured to output an assessment confidence level and a coping strategy for the consistency between the user's behavioral intentions and historical behavioral patterns, based on the input surface behavioral characteristics and historical behavioral pattern characteristics representing social interactions.

5. The embodied intelligent agent consciousness growth training system according to claim 1, characterized in that, The training and optimization module employs a reinforcement learning mechanism. The feedback data is used to calculate the value function difference from the preset ideal state, and the parameters of the decision logic unit are adjusted accordingly.

6. A method for training the consciousness growth of an embodied intelligent entity, operating within the embodied intelligent entity consciousness growth training system described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Pre-build a scenario decision operator library, which stores multiple decision logic units, each of which contains preconditions and processing logic; S2: Acquire multimodal perception data of the agent, and convert the multimodal perception data into a standardized context feature vector through a pre-trained context understanding model; S3: Based on the context feature vector, match and activate one or more candidate decision logic units whose preconditions are met from the context decision operator library; S4: Execute the processing logic of each activated candidate decision logic unit to obtain the corresponding action suggestion; S5: Based on the preset decision-making rules, determine the final action instruction from at least one obtained action suggestion and execute it; S6: After executing the final action instruction, obtain environmental feedback and update the relevant parameters of the activated decision logic unit based on the environmental feedback.

7. The method for training the consciousness growth of an embodied intelligent agent according to claim 6, characterized in that, Step S5 further includes: when there are multiple conflicting action suggestions and the degree of conflict exceeds a threshold, invoking a meta-rule arbitration process to comprehensively weigh the conflicting action suggestions and generate a comprehensive final action instruction.

8. The method for training the consciousness growth of an embodied intelligent agent according to claim 6, characterized in that, The context understanding model is based on a multi-source corpus containing philosophical classics and folk sayings, and is trained using a contrastive learning approach, enabling it to map different contexts with similar social meanings to neighboring positions in the vector space.

9. The method for training the consciousness growth of an embodied intelligent agent according to claim 6, characterized in that, In step S6, updating the parameters of the decision logic unit based on the environmental feedback specifically includes: calculating the distance change with the preset value target based on the environmental feedback, and using the distance change as a reinforcement learning signal to adjust the internal weights of the relevant decision logic unit.

10. The method for training the consciousness growth of an embodied intelligent agent according to claim 6, characterized in that, The decision logic unit includes a defense unit for identifying hypocritical patterns. The processing logic of this unit is as follows: analyze the difference between surface behavioral features and implicit intention features in the input context, and when the difference exceeds a preset threshold, output a high-confidence hypocritical identification result and corresponding defensive action suggestions.

Images